Lossless Image Coding Technology

With the rapid development of communication technology and explosion ofinformation, digital image technology is widely used in all areas of life. However, thelarge amount of data is generated when information is converted into a digital image.This is why highly efficient compression technology becomes an urgent need. Manyapplications such as medicine and satellite require very high quality of images.Therefore, study of lossless compression technology is of great significance. Losslessimage compression mainly includes mapping and the entropy coding, and CALICknown to have better compression effect is a lossless coding method. In CALIC,arithmetic entropy coding has been adopted and a non-linear gradient predictionmethod has been proposed. Pixels are estimated by weighting adjacent pixels throughgradient prediction in this nonlinear prediction. Compared with the previous methodof linear prediction, this nonlinear prediction enhances the edge pixel accuracy.CALIC has high compression efficiency by carefully designed context model andarithmetic coding. However, due to the singularity of the prediction scan mode, theprediction reference pixel range of options is limited and the prediction accuracy isaffected.This thesis mainly focuses on arithmetic entropy coding and prediction ofCALIC. It has the following main contributions.First of all, this thesis proposes the arithmetic coding data distributioncharacteristics and a better implementation, through analysis of characteristics ofarithmetic coding principle and comparison of multiple existing implementations. In avariety of entropy coding algorithms, arithmetic coding, approximating the idealentropy in theory, has received extensive attention. But as a result of calculationaccuracy problem, there is a gap between the theoretical and actual compressionresults. In data analysis, experiment shows that the arithmetic coding can adaptivelyupdate the probability and the compression efficiency of the arithmetic coding onlyfluctuates within a certain range when the data distribution is too extreme. Thecomparative experiment proves that improved method with as approximation isrelatively optimal and its relative efficiency loss is about2%.Secondly, for the single raster scan problem in gradient prediction method ofCALIC, this thesis proposes a new prediction method which includes horizontal,vertical, left-oblique and right-oblique four scan mode gradient prediction. Eachcoding block is matched with different scanning based on the texture properties in order to exploit the redundancy within the image fully and provide more relevantreference pixels. Thus, according to the specific mode of gradient value, adjacentpixels are weighted as the predicted value of the current pixel. Experiment shows thatthe improved prediction method can effectively improve the prediction accuracy andreduce the prediction residual zero-order entropy about3%to12%.In this thesis, we also propose two methods on a variety of scan mode selection.One is exhaustive method. Each coding block adopts four different scanning modegradient prediction methods and selects the optimal scan mode based on the sum ofsquares of cumulative predicted residual. The other one is adaptive selection method.Refer to the sample code block images to calculate the direction of the gradient value.Forecast local texture property and then select the appropriate scan mode. Experimentshows that the results with the adaptive selection method mode basic approximate theones with exhaustive method in the selection mode and the prediction residualzero-order entropy.